Relation Extraction using Explicit Context Conditioning

Proceedings of NAACL-HLT 2019, pages 1442–1447 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1442 Relation Extraction using Explicit Context Conditioning Gaurav Singh∗ University College London g.singh@cs.ucl.ac.uk Parminder Bhatia Amazon parmib@amazon.com Abstract Relation Extraction (RE) aims to label rela- tions between groups of marked entities in raw text. Most current RE models learn context-aware representations of the target en- tities that are then used to establish relation between them. This works well for intra- sentence RE and we call them first-order re- lations. However, this methodology can some- times fail to capture complex and long depen- dencies. To address this, we hypothesize that at times two target entities can be explicitly connected via a context token. We refer to such indirect relations as second-order rela- tions and describe an efficient implementation for computing them. These second-order rela- tion scores are then combined with first-order relation scores. Our empirical results show that the proposed method leads to state-of-the- art performance over two biomedical datasets. 1 Introduction There are wide applications for Information Ex- traction in general (Jin et al., 2018) and Relation Extraction (RE) in particular, one reason why re- lation extraction continues to be an active area of research (Bach and Badaskar, 2007; Kambhatla, 2004; Kumar, 2017). Traditionally, a standard RE model would start with entity recognition and then pass the extracted entities as inputs to a separate relation extraction model, which meant that the er- rors in entity recognition were propagated to RE. This problem was addressed by end-to-end mod- els (Miwa and Bansal, 2016; Zheng et al., 2017; Adel and Sch¨utze, 2017; Bhatia et al., 2018) that jointly learn both NER and RE. Generally, these models consist of an encoder followed by a relationship classification (RC) unit (Verga et al., 2018; Christopoulou et al., 2018; ∗G. Singh was an intern at Amazon at the time of work College France Bill ........... located_in located_in located_in Figure 1: Pictorial representation of a second-order re- lation between two entities (Bill & France) connected by a context token (College). Su et al., 2018). The encoder provides context- aware vector representations for both target enti- ties, which are then merged or concatenated be- fore being passed to the relation classification unit, where a two layered neural network or multi- layered perceptron classifies the pair into different relation types. Such RE models rely on the encoder to learn ‘perfect’ context-aware entity representations that can capture complex dependencies in the text. This works well for intra-sentence relation extrac- tion i.e. the task of extracting relation from enti- ties contained in a sentence (Christopoulou et al., 2018; Su et al., 2018). As these entities are closer together, the encoder can more easily establish connection based on the language used in the sen- tence. Additionally, these intra-sentence RE mod- els can use linguistic/syntactical features for an improved performance e.g. shortest dependency path. Unfortunately, success in intra-sentence RE has not been replicated for cross-sentence RE. As an example, a recent RE method called …

Content truncated to 500 words